GRASP: GRaph-Augmented Score Propagation for OOD detection

This is the source code for NeurIPS 2024 submitted paper Revisiting Score Propagation in Graph Out-of-Distribution Detection.

Dependence

• Ubuntu 20.04.6
• Cuda 11.3
• Pytorch 1.13.1
• Pytorch Geometric 2.3.1

Dataset Preparation

• All small-scale datasets are already in the codebase or will be downloaded automatically during data loading.

• We upload all large-scale datasets except wiki to google drive folder. Download these files and put these datasets to data foler.

• For dataset wiki, please download wiki_features.pt, wiki_edges.pt from the links. Then put the downloaded files to the ./data/ directory.

Usage

1. Post-hoc Methods

All pretrained model checkpoints are already in the codebase.

Run ./scripts/test_ood.sh to evaluate the performance of all post-hoc methods.

2. Training-based Methods

The folder ood_training contains all codes for training-based methods GKDE, GPN and OODGAT. Enter folder ood_training and run scripts/run.sh to evaluate these methods.